A Personalized, Non-Invasive Blood Glucose Prediction System Using a CNN-LSTM Model with a Sim-to-Real Transfer Learning Strategy

The translation of Non-Invasive Glucose Monitoring (NIBGM) into routine clinical care is currently hindered by two significant engineering hurdles: the inherently weak signal fidelity, characterized by a low Signal-to-Noise Ratio (SNR), in optical sensors, and the "Cold Start" constraint common in deep learning applications. Traditional recursive architectures, such as Long Short-Term Memory networks (LSTMs), often struggle to generalize in data-sparse contexts typical of personalized medicine. To address these limitations, this study presents the Hybrid Temporal-Attention Network (HTAN), a novel framework that disentangles sensor artifacts from metabolic trends by integrating Residual Convolutional Neural Networks (CNNs) with Bi-directional LSTMs and Multi-Head Attention mechanisms. The validity of this architecture was assessed using a "Dual-Stream" protocol, with pre-training on a high-precision "Chaos-Augmented" Digital Twin ( samples) followed by benchmarking against three distinct real-world clinical cohorts ( ). Our analysis reveals a significant performance divergence by phenotype: while conventional Random Forest models achieve adequate accuracy for stable Type 2 Diabetes (MARD = 6.18%), the HTAN model exhibits superior robustness in high-volatility Type 1 Diabetes scenarios. Notably, on the complex Shanghai T1DM dataset (N=7), standard Bi-LSTMs struggled to track the dynamics effectively (Mean Absolute Relative Difference, MARD = 8.65%), whereas HTAN achieved a markedly lower MARD of 5.11%. These results indicate that attention-based methodologies, when initialized with synthetic physiological priors, offer a viable approach to modelling intricate metabolic instability in regimes with limited data.